Device, configured to operate a machine learning system based on predefinable rollout

Abstract

A device for operating a machine learning system. The machine learning system is assigned a predefinable rollout, which characterizes a sequence in which each of the layers ascertains an intermediate variable. When assigning the rollout, each connection or each layer is assigned a control variable, which characterizes whether the intermediate variable of each of the subsequent connected layers is ascertained according to the sequence or regardless of the sequence. A calculation of an output variable of the machine learning system as a function of an input variable of the machine learning system is controlled as a function of the predefinable rollout. Also described is a method for operating the machine learning system.

What is claimed is: 
     
       1. A device configured to operate a machine learning system, the machine learning system including a plurality of layers, which are connected with the aid of connections, the device comprising:
 a machine-readable memory element, on which commands are stored which, when executed by a computer, ensure that the computer carries out a method that includes the following steps:
 assigning to the machine learning system a predefinable rollout, which characterizes a sequence, according to which the layers each ascertain an intermediate variable, 
 when assigning the predefinable rollout, assigning to each connection or each layer a control variable, which characterizes whether the intermediate variable of the respective subsequent connected layers is ascertained according to the sequence or regardless of the sequence, and 
 calculating an output variable of the machine learning system as a function of an input variable of the machine learning system, the calculating being controlled as a function of the predefinable rollout, 
 wherein the control variables of the rollout are selected at random or as a function of an additional predefinable rollout. 
 
 
     
     
       2. The device as recited in  claim 1 ,
 wherein when controlling the calculation, each of the layers that ascertains according to the sequence ascertains step-wise and in succession the intermediate variable according to the sequence of the rollout, in each case at a predefinable point in time of a sequence of points in time, and those layers that ascertain their intermediate variables regardless of the sequence each ascertain their intermediate variables, in each case at each step at the respective predefinable points in time, and 
 wherein the machine learning system is assigned a plurality of different rollouts, in each case the calculating of the machine learning system being controlled as a function of the assigned rollouts, the controlled calculating of the machine learning system for each of the assigned rollouts being compared with one another based on at least one predefinable comparison criterion, the predefinable rollout being selected as a function of the comparison. 
 
     
     
       3. The device as recited in  claim 1 , wherein the sequence is executed step by step, in each step at least one of the layers ascertaining its output variable according to the sequence. 
     
     
       4. The device as recited in  claim 1 , wherein the machine learning system includes at least one skip connection, which connects a first layer to a second layer and the first layer and the second layer are also directly connected with the aid of at least two connections. 
     
     
       5. The device as recited in  claim 4 , wherein when assigning the rollout, those connections that connect a first layer with a second layer and the first layer and the second layer are also directly connected with the aid of at least two connections, are assigned the control variable, so that the intermediate variable of the second layer is ascertained regardless of the sequence. 
     
     
       6. The device as recited in  claim 1 , wherein the machine learning system includes at least one recurrent connection. 
     
     
       7. The device as recited in  claim 1 , wherein those layers that ascertain their intermediate variables regardless of the sequence, ascertain their intermediate variables as a function of a chronologically preceding intermediate variable of a chronologically preceding calculation step of the previous layer, and those layers that ascertain their intermediate variable according to the sequence, ascertain their intermediate variable as a function of a chronologically instantaneous intermediate variable of an instantaneous calculation step of the preceding layer. 
     
     
       8. The device as recited in  claim 1 , wherein the machine learning system does not include a closed path. 
     
     
       9. The device as recited in  claim 1 , wherein the intermediate variables of those layers, that ascertain their intermediate variable regardless of the sequence, are each ascertained in parallel. 
     
     
       10. The device as recited in  claim 9 , wherein the ascertainment in parallel of the intermediate values is carried out on processor cores connected in parallel. 
     
     
       11. The device as recited in  claim 1 , wherein the intermediate variables of those layers, that ascertain their intermediate variable regardless of the sequence, are ascertained asynchronously. 
     
     
       12. The device as recited in  claim 1 , wherein after the machine learning system is provided an input variable for the first time, it is checked after each step during a step-wise ascertainment according to the sequence of the output variable of the machine learning system, whether those layers that ascertain their intermediate variable regardless of the sequence are each provided an already ascertained intermediate variable of a previous layer. 
     
     
       13. The device as recited in  claim 1 , wherein a plurality of the control variables of the predefinable rollout characterize that respective intermediate variables are ascertained regardless of the sequence. 
     
     
       14. The device as recited in  claim 1 , wherein during the calculation of the machine learning system, the machine learning system is provided a sequence of input variables of an input layer of the machine learning system, in direct succession, in each case, at a time step of a sequence of time steps, each layer ascertaining at each time step as a function of an input variable, the respective intermediate variable, which in each case is assigned to one of the input variables. 
     
     
       15. The device as recited in  claim 1 , wherein in the case of one of the rollouts, all connections and layers are each assigned the same control variable, so that each of the output variables is ascertained regardless of the sequence. 
     
     
       16. The device as recited in  claim 1 , wherein in the case of one of the rollouts, all connections or layers are each assigned the same control variable, so that each of the output variables is ascertained regardless of the sequence. 
     
     
       17. The device as recited in  claim 1 , wherein the rollouts are compared with one another based on the predefinable comparison criterion, the predefinable criterion as a function of the control of the machine learning system being ascertained as a function of the respectively assigned rollout, the predefinable criterion including:
 a first variable, which characterizes a number of time steps required in order, starting with a first time step at which the input layer is provided the input variable, to ascertain the output variable up to a second time step, the output layer not being connected to any additional layer. 
 
     
     
       18. The device as recited in  claim 17 , wherein the predefinable criterion includes a variable that characterizes how many output variables the machine learning system ascertains within a predefinable number of time steps. 
     
     
       19. The device as recited in  claim 17 , wherein the predefinable criterion includes a variable that characterizes how reliable an accuracy of the output variable of the machine learning system is with the aid of the respective rollout. 
     
     
       20. The device as recited in  claim 17 , wherein the predefinable criterion includes a variable that characterizes a period of time after which a start-up phase is completed, or the classification accuracy has reached a maximum value. 
     
     
       21. The device as recited in  claim 17 , wherein the predefinable criterion includes a variable that characterizes how many connections in direct succession include the same control variable. 
     
     
       22. The device as recited in  claim 1 , wherein the rollouts are also compared with a rollout based on the predefinable comparison criterion, in which all control variables provide the processing of the results according to the sequence. 
     
     
       23. The device as recited in  claim 1 , wherein the layers of the machine learning system are in each case a layer of a deep neuronal network. 
     
     
       24. The device as recited in  claim 23 , wherein the machine learning system classifies an image sequence. 
     
     
       25. The device as recited in  claim 24 , wherein the classification that takes place image element-wise is segmented. 
     
     
       26. The device as recited in  claim 1 , wherein the input variable of the input layer is a detected sensor variable and a control variable is ascertained as a function of the calculation of the machine learning system. 
     
     
       27. The device as recited in  claim 1 , wherein the device is used for training the machine learning system. 
     
     
       28. The device as recited in  claim 1 , wherein the device is used for a real time processing of a video with the aid of the machine learning system. 
     
     
       29. The device as recited in  claim 1 , wherein the device is for controlling a calculation of the machine learning system. 
     
     
       30. A device configured to operate a machine learning system, the machine learning system including a plurality of layers, which are connected with the aid of connections, the device comprising:
 a machine-readable memory element, on which commands are stored which, when executed by a computer, ensure that the computer carries out a method that includes the following steps:
 assigning to the machine learning system a predefinable rollout, which characterizes a sequence, according to which the layers each ascertain an intermediate variable, 
 when assigning the predefinable rollout, assigning to each connection or each layer a control variable, which characterizes whether the intermediate variable of the respective subsequent connected layers is ascertained according to the sequence or regardless of the sequence, and 
 calculating an output variable of the machine learning system as a function of an input variable of the machine learning system, the calculating being controlled as a function of the predefinable rollout, 
 wherein at least one of the control variables of the predefinable rollout is changed as a function of a disruption of the calculation of the machine learning system. 
 
 
     
     
       31. The device as recited in  claim 30 , wherein the disruption includes that the machine learning system has erroneously ascertained the output variable or one of the intermediate variables.